https://gcc.gnu.org/bugzilla/show_bug.cgi?id=85358
Bug ID: 85358
Summary: PowerPC: Using -mabi=ieeelongdouble -mcpu=power9
breaks __ibm128
Product: gcc
Version: 8.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: meissner at gcc dot gnu.org
Target Milestone: ---
If you switch the default long double type to IEEE 128-bit
(-mabi=ieeelongdouble), the GCC compiler will crash.
Using this code:
__ibm128 add (__ibm128 a, __ibm128 b) { return a+b; }
This error is generated:
during RTL pass: expand
foo02.c: In function ‘add’:
foo02.c:1:49: internal compiler error: Segmentation fault
__ibm128 add (__ibm128 a, __ibm128 b) { return a+b; }
~^~
0xb683ff crash_signal
/home/meissner/fsf-src/pr85075/gcc/toplev.c:325
0x86bd0f aggregate_value_p(tree_node const*, tree_node const*)
/home/meissner/fsf-src/pr85075/gcc/function.c:2040
0x6d76b4 emit_library_call_value_1(int, rtx_def*, rtx_def*, libcall_type,
machine_mode, int, std::pair<rtx_def*, machine_mode>*)
/home/meissner/fsf-src/pr85075/gcc/calls.c:4771
0xf2ae05 emit_library_call_value(rtx_def*, rtx_def*, libcall_type,
machine_mode, rtx_def*, machine_mode)
/home/meissner/fsf-src/pr85075/gcc/rtl.h:4156
0xf01ed0 rs6000_expand_float128_convert(rtx_def*, rtx_def*, bool)
/home/meissner/fsf-src/pr85075/gcc/config/rs6000/rs6000.c:22489
0x10f318d gen_extenddfkf2(rtx_def*, rtx_def*)
/home/meissner/fsf-src/pr85075/gcc/config/rs6000/rs6000.md:7553
0xa608db maybe_emit_unop_insn(insn_code, rtx_def*, rtx_def*, rtx_code)
/home/meissner/fsf-src/pr85075/gcc/optabs.c:3547
0xa60988 emit_unop_insn(insn_code, rtx_def*, rtx_def*, rtx_code)
/home/meissner/fsf-src/pr85075/gcc/optabs.c:3570
0x80cc05 convert_mode_scalar
/home/meissner/fsf-src/pr85075/gcc/expr.c:601
0x80cc05 convert_move(rtx_def*, rtx_def*, int)
/home/meissner/fsf-src/pr85075/gcc/expr.c:268
0x80d891 convert_modes(machine_mode, machine_mode, rtx_def*, int)
/home/meissner/fsf-src/pr85075/gcc/expr.c:709
0xa6451f expand_binop(machine_mode, optab_tag, rtx_def*, rtx_def*, rtx_def*,
int, optab_methods)
/home/meissner/fsf-src/pr85075/gcc/optabs.c:1317
0x81b359 expand_expr_real_2(separate_ops*, rtx_def*, machine_mode,
expand_modifier)
/home/meissner/fsf-src/pr85075/gcc/expr.c:9715
0x808b53 expand_expr_real_1(tree_node*, rtx_def*, machine_mode,
expand_modifier, rtx_def**, bool)
/home/meissner/fsf-src/pr85075/gcc/expr.c:9918
0x6e8abd expand_expr
/home/meissner/fsf-src/pr85075/gcc/expr.h:276
0x6e8abd expand_return
/home/meissner/fsf-src/pr85075/gcc/cfgexpand.c:3573
0x6e8abd expand_gimple_stmt_1
/home/meissner/fsf-src/pr85075/gcc/cfgexpand.c:3663
0x6e8abd expand_gimple_stmt
/home/meissner/fsf-src/pr85075/gcc/cfgexpand.c:3790
0x6ea667 expand_gimple_basic_block
/home/meissner/fsf-src/pr85075/gcc/cfgexpand.c:5819
0x6efdce execute
/home/meissner/fsf-src/pr85075/gcc/cfgexpand.c:6425
Please submit a full bug report,
with preprocessed source if appropriate.
Please include the complete backtrace with any bug report.
See <https://gcc.gnu.org/bugs/> for instructions.
Now what is happening is since there is hardware support for IEEE 128-bit
operations in power9, the GCC compiler decides it can do the calculation of IBM
double double in IEEE 128-bit and tries to do the conversion.
However, rather than converting to TFmode (long double), it converts to KFmode
(__float128/_Float128). However, KFmode isn't properly registered (since
TFmode is used for float128), and emit_library_call_value_1 gets a NULL tree
type which aggregate_value_p crashes with a NULL pointer.
Ideally, for __ibm128 we should NOT try to use __float128 for the arithmetic,
since there are cases where the last bit is not accurate.
